From Garbage to Gold rethinks GIGO

Strong theory paper, but its reach is narrower than the headline suggests: the result depends on latent hierarchical structure, not on “dirty data” in general.

• –The clean split between predictor error and structural uncertainty is the paper’s most useful idea; it explains why more variables can help when they are distinct proxies for shared latent causes.
• –The breadth-over-depth result is especially persuasive for EHR and warehouse-style tabular systems, where feature redundancy can capture hidden signal better than exhaustive cleaning of a small set.
• –The benign overfitting connection is a nice bridge to existing literature, especially if the proxy features really induce low-rank-plus-diagonal covariance.
• –The clinical case study makes the theory feel grounded, but it should be read as motivation rather than proof.
• –The biggest practical caveat is assumption sensitivity: if the latent hierarchy is weak or absent, this is not a substitute for careful curation.